SeACo-Paraformer - A non-autoregressive automatic speech recognition system with hotword customization for enhanced Chinese speech recognition.

## Overview of SeACo-Paraformer SeACo-Paraformer is a non-autoregressive end-to-end automatic speech recognition (ASR) system designed specifically for Chinese speech recognition. It integrates hotword customization to improve the accuracy of recognizing specific entity words such as names and places. ## Key Features of SeACo-Paraformer The key features of SeACo-Paraformer include: 1. Non-autoregressive (NAR) design: This architecture provides higher efficiency during the inference phase compared to traditional autoregressive models. 2. Hotword customization: Users can specify hotwords (e.g., names, entity words) to improve their recognition accuracy. 3. Combination of accuracy and efficiency: It combines the accuracy of attention encoder-decoder (AED) models with the efficiency of NAR models, ensuring a balance between performance and speed. 4. Explicit customization capacity: The system handles hotword customization in an explicit manner, providing superior customization performance. ## Functions of SeACo-Paraformer The main functions of SeACo-Paraformer include: - Automatic Speech Recognition (ASR): Converts speech signals into text, suitable for general Chinese scenarios. - Hotword customization: Allows users to specify hotwords during the ASR process to enhance the recognition accuracy of specific words. ## Access and Usage of SeACo-Paraformer SeACo-Paraformer can be accessed and used via the FunASR toolkit, an open-source speech recognition toolkit. Users can: - Download pre-trained models from platforms like ModelScope or Hugging Face. - Use the FunASR API for model inference, especially in scenarios requiring hotword customization. Detailed usage instructions can be found in the FunASR GitHub repository and related documentation. ## Datasets Used for Testing SeACo-Paraformer SeACo-Paraformer was tested on a large-scale industrial dataset consisting of 50,000 hours of data. It demonstrated superior performance in both hotword customization and general ASR tasks. ## Additional Resources for SeACo-Paraformer More information about SeACo-Paraformer can be found at the following URLs: - [ModelScope Page](https://modelscope.cn/models/iic/speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary) - [SeACo-Paraformer GitHub Repository](https://github.com/R1ckShi/SeACo-Paraformer) - [FunASR GitHub Repository](https://github.com/modelscope/FunASR) - [SeACo-Paraformer Paper on arXiv](https://arxiv.org/abs/2308.03266) ### Citation sources: - [SeACo-Paraformer](https://modelscope.cn/models/iic/speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary) - Official URL Updated: 2025-03-26